Scribd
Upload
26 views

Least-Square Method

The document provides an overview of the least-squares method for fitting models to data. It discusses: 1) How least squares is used to find the "next best thing" to an exact solution when solving inconsistent systems of equations, or to approximately fit data when an exact fit is not possible. 2) How to derive the normal equations for a least squares problem in matrix form as AT(Ax - b) = 0 and solve them to find the least squares solution. 3) Examples of using least squares to find the best fitting line and parabola to different data sets. The residual errors and RMSE are calculated to evaluate the fit. 4) The general least squares procedure of choosing a

Uploaded by

Santiago Sarmiento
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
26 views

Least-Square Method

The document provides an overview of the least-squares method for fitting models to data. It discusses: 1) How least squares is used to find the "next best thing" to an exact solution when solving inconsistent systems of equations, or to approximately fit data when an exact fit is not possible. 2) How to derive the normal equations for a least squares problem in matrix form as AT(Ax - b) = 0 and solve them to find the least squares solution. 3) Examples of using least squares to find the best fitting line and parabola to different data sets. The residual errors and RMSE are calculated to evaluate the fit. 4) The general least squares procedure of choosing a

Uploaded by

Santiago Sarmiento
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 32

Least-Square Method

Least-Square basics

Josept Revuelta Ph.D.


Introduction
The need for least squares methods comes from two different directions, we
know now how to find the solution of Ax = b when a solution exists. However,
we do not know what to do when there is no solution. When the equations are
inconsistent, which is likely if the number of equations exceeds the number of
unknowns, the answer is to find the next best thing: the least squares
approximation.
Additionally, we found polynomials that exactly fit data points. However, if the
data points are numerous, or the data points are collected only within some
margin of error, fitting a high-degree polynomial exactly is rarely the best
approach. In such cases, it is more reasonable to fit a simpler model that may
only approximate the data points. Both problems, solving inconsistent systems
of equations and fitting data approximately, are driving forces behind least
squares.
Inconsistent systems of equations
It is not hard to write down a system of equations that has no solutions.
Consider the following three equations in two unknowns:

Any solution must satisfy the first and third equations, which cannot both
be true. A system of equations with no solution is called inconsistent. What
is the meaning of a system with no solutions? Perhaps the coefficients are
slightly inaccurate. In many cases, the number of equations is greater than
the number of unknown variables, making it unlikely that a solution can
satisfy all the equations.
If we choose this “closeness’’ to mean close in Euclidean distance, there is a
straightforward algorithm for finding the closest 𝑥.ҧ This special 𝑥ҧ will be
called the least squares solution. We can get a better picture of the failure
of system to have a solution by writing it in a different way. The matrix
form of the system is A𝑥ҧ = b, or

The alternative view of matrix/vector multiplication is to write the equivalent


equation
In fact, any m x n system A𝑥ҧ = b can be viewed as a vector equation

which expresses b as a linear combination of the columns vi of A, with


coefficients x1, . . . , xn. In our case, we are trying to hit the target vector b as a
linear combination of two other three-dimensional vectors.
Important background

First, we establish some notation. Recall the concept of the transpose AT of


the m × n matrix A, which is the n × m matrix whose rows are the columns of
A and whose columns are the rows of A, in the same order. The transpose of
the sum of two matrices is the sum of the transposes, (A + B)T = AT + BT . The
transpose of a product of two matrices is the product of the transposes in the
reverse order, that is, (AB)T = BT AT .

To work with perpendicularity, recall that two vectors are at right angles to
one another if their dot product is zero. For two m-dimensional column
vectors u and v, we can write the dot product solely in terms of matrix
multiplication by
The vectors u and v are perpendicular, or orthogonal, if uT · v = 0, using ordinary
matrix multiplication.

Now we return to our search for a formula for 𝑥.ҧ We have established that

Expressing the perpendicularity in terms of matrix multiplication, we find


that
Using the preceding fact about transposes, we can rewrite this expression as

meaning that the n-dimensional vector AT(b − Ax) is perpendicular to every


vector 𝑥ҧ in Rn, including itself. There is only one way for that to happen:

This gives a system of equations that defines the least squares solution,

This system of equations is known as the normal equations. Its solution 𝑥ҧ is


the so-called least squares solution of the system A𝑥ҧ = b.
Example 1
Use the normal equations to find the least squares solution of the inconsistent
system. The problem in matrix form A𝑥ҧ = b has
The components of the normal equations are

and
The normal equations

can now be solved by Gaussian elimination. The tableau form is

which can be solved to get x = (x1,x2) = (7/4,3/4).


Substituting the least squares solution into the original problem yields

To measure our success at fitting the data, we calculate the residual of the
least-squares solution 𝑥ҧ as
If the residual is the zero vector, then we have solved the original system
A𝑥ҧ = b exactly. If not, the Euclidean length of the residual vector is a
backward error measure of how far 𝑥ҧ is from being a solution.

There are at least three ways to express the size of the residual. The Euclidean
length of a vector,

is a norm in the sense, called the 2-norm. The squared error


and the root mean squared error (the root of the mean of the squared error)

are also used to measure the error of the least-squares solution. The three
expressions are closely related; namely
so finding the x that minimizes one, minimizes all. For Example 1, the SE =
(.5)2 + 02 + (−.5)2 = 0.5, the 2-norm of the error is ||r||2 = 0.5 ≈ 0.707, and the
RMSE = 0.5/3 = 1/ 6 ≈ 0.408.
Example 2
Solve the least squares problem:

The normal equations AT Ax = AT b are

The solution of the normal equations are x1 = 3.8 and x2 = 1.8.


The residual vector is

which has Euclidean norm


Fitting model data
Let (t1,y1), . . . , (tm,ym) be a set of points in the plane, which we will often refer
to as the “data points.’’ Given a model, such as all lines y = c1 + c2t , we can
seek to locate the specific instance of the model that best fits the data points in
the 2-norm. The the least-squares idea consists of measuring the residual of
the fit by the squared errors of the model at the data points and finding the
model parameters that minimize this quantity. This criterion is displayed in
the following figure:
Example 3
Find the line that best fits the three data points (t ,y) = (1,2), (−1,1), and (1,3)
in the figure below
The model is y = c1 + c2t , and the goal is to find the best c1 and c2. Substitution
of the data points into the model yields

or, in matrix form,


We know this system has no solution (c1,c2) for two separate reasons. First,
if there is a solution, then the y = c1 + c2t would be a line containing the
three data points. However, it is easily seen that the points are not
collinear. Second, this is the system of equation that we discussed at the
beginning of this class. We noticed then that the first and third equations
are inconsistent, and we found that the best solution in terms of least
squares is (c1,c2) = (7/4,3/4). Therefore, the best line is y = 7/4 + 3/4t.

We can evaluate the fit by using the statistics defined earlier. The residuals at
the data points are

and the RMSE is 1/ 6, as seen earlier.


Fitting data by least squares
Given a set of m data points (t1,y1), . . . , (tm,ym).

STEP 1. Choose a model. Identify a parameterized model, such as y =


c1 + c2t , which will be used to fit the data.

STEP 2. Force the model to fit the data. Substitute the data points into
the model. Each data point creates an equation whose unknowns are
the parameters, such as c1 and c2 in the line model. This results in a
system Ax = b, where the unknown x represents the unknown
parameters.

STEP 3. Solve the normal equations. The least squares solution for the
parameters will be found as the solution to the system of normal
equations AT Ax = AT b.
Example 4

Find the best line and best parabola for the four data points (−1,1), (0,0),
(1,0), (2,−2) in the figure below
In accordance with the preceding procedure, we will follow three steps:
(1) Choose the model y = c1 + c2t as before. (2) Forcing the model to fit the data
yields

or, in matrix form,


(3) The normal equations are

Solving for the coefficients c1 and c2 results in the best line y = c1 + c2t = 0.2 −
0.9t. The residuals are
The error statistics are
Next, we extend this example by keeping the same four data points but
changing the model. Set y = c1 + c2t + c3t2 and substitute the data points to yield

or, in matrix form,


This time, the normal equations are three equations in three unknowns:

Solving for the coefficients results in the best parabola y = c1 + c2t + c3t2 =
0.45 − 0.65t − 0.25t2. The residual errors are given in the following table:

The error statistics are


The Matlab commands polyfit and polyval are designed not only to interpolate
data, but also to fit data with polynomial models. For n input data points,
polyfit used with input degree n − 1 returns the coefficients of the interpolating
polynomial of degree n − 1. If the input degree is less than n − 1, polyfit will
instead find the best least squares polynomial of that degree. For example, the
commands
Exercises
Find the best parabola through each data point set in the previous exercise and compare
RMSE with the best line fit.

You might also like